CNN-based Image Denoising for Outdoor Active Stereo
Stereo vision has been the most widely used passive 3D sensing technology for a variety of vision tasks. 3D coordinates are computed by triangulating correspondences found in the stereo image pair. For homogeneous areas where stereo matching fails, a stereo projector system can be employed by actively projecting auxiliary texture onto the scene. However, the applicability of this approach is restricted to indoor scenarios, since in outdoor environment where the sunlight is strong, the projected pattern is almost invisible. A simple increase in contrast of the projection leads to dramatic rise of the noise level, which again has an adverse impact on the matching algorithm. We propose a novel framework to tackle this problem, exploiting adaptive contrast improvement with denoising techniques using convolutional neural networks (CNNs) on the difference images to digitally enhance the projection, which is later added back onto the image pair to assist stereo matching. In order to learn an optimal denoising network dedicated to the projected pattern, a straightforward workflow is devised to allow for convenient acquisition of noisy and noiseless pattern images for the input and ground truth respectively. Extensive evaluation on real-world data compared to the state of the art justifies the effectiveness of not only the presented denoising CNN architecture and training routine, but also the entire pipeline for outdoor active stereo reconstruction.